CORAL reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

The four reports share the frozen boundary and evidence limits. They distinguish implementation and doctrine/design from observed operation and causal benefit. This reconciliation uses their pinned records and passages only; it adds no source inspection. No correction answers or declined objections were supplied.

Identity and containment

No duplicate supersession is made. The overlaps below are valid containers, narrower mechanisms or distinct consumers, rather than interchangeable whole-record identities. Keeping them does not combine their statuses.

  • MEM-OBJ-raw-sources covers raw Markdown captures and review sidecars. EPI-OBJ-raw-sources isolates the captured source assertions; EPI-OBJ-review-json isolates the requested per-claim review output. The shared deep-research capture passage identifies the same raw-file surface, but the epistemic records separate input evidence from judgments about it. Whole-record identity is ruled out by the memory container's additional sidecar scope. Its findings remain valid when these already-declared parts are separated, so the container is not superseded. Neither record establishes an actual capture or review.
  • MEM-OBJ-notes includes research, experiment, infrastructure, focus and synthesis notes, metadata and legacy formats. EPI-OBJ-research-notes addresses research, experiment and synthesis claims with their status/confidence/lineage metadata. This is overlapping content within the broader memory inventory, not identity of the whole referents. MEM-OBJ-note-navigation separately owns index and coverage-ledger surfaces; a research writer's use of those surfaces does not make them another research-note declaration. No combined finding is invalidated by retaining these distinctions.
  • MEM-RTE-knowledge-revision spans research and result-fed notes, roles and skills, including warm-start and later consumption. EPI-RTE-research-writing traces source acquisition and note generation; EPI-RTE-synthesis-review traces the reviewer; EPI-RTE-confidence-disposition traces the author's ensuing reliance decision. These are narrower functions, not duplicate whole revision routes. The memory route retains its afforded implementation finding; the epistemic routes retain their claimed findings and doctrine-only ledger assessments. Neither report claims completed synthesis, an executed review or an observed acceptance decision.
  • MEM-RTE-knowledge-query spans CLI, API and direct filesystem delivery for notes and skills. EPI-RTE-note-selection isolates hub status selection and graph projection. The latter is not the whole query route. Its requested status predicate does not contradict the memory finding that numeric CLI reads omit the status filter: those are different caller paths.
  • MEM-CLM-grounding-gate addresses structured-trace verification language, reviewer requirements, grounding and lint limits. EPI-CLM-grounding-check isolates the research skill's mechanical-enforcement claim. Their scope is not identical. EPI-RTE-grounding-check and EPI-RTE-note-lint declare the separate evaluators, not further copies of those claims.
  • WF-RTE-candidate-export isolates export from the broader direct revert/checkout/export family in EPI-RTE-code-selection. Their export findings converge on caller-selected branch creation without a score gate, but the broader record also covers active-code changes. Neither valid route is superseded by the other.

Explicit containment remains intact: EPI-OBJ-attempt-score and EPI-OBJ-attempt-feedback are parts of RT-OBJ-attempt, with different forms and consumers. MEM-OBJ-migration-arrival is a retained output part of RT-RTE-island-migration, not another migration route. MEM-OBJ-checkpoint is the separate shared-state repository linked by the attempt, not the attempt itself. These containers require no replacement.

Connected routes and independent convergence

Launch and delivery of guidance are not completion of knowledge work. RT-RTE-agent-launch, RT-BAP-generated-instructions and RT-RTE-heartbeat-feedback supply files, prompts and process transitions. MEM-RTE-knowledge-revision concerns the model-authored retained changes those prompts request. MEM-BAP-retained-guidance concerns the later authority of that mutable material, not the authority of the generated instructions that point to it. These are consecutive edges, not duplicate identities. The memory integration request not to merge synthesis with heartbeat transport is retained.

Grading and explanatory feedback have different boundaries. RT-RTE-grader-execution supports task-defined computation and finalized results. The epistemic ledger uses that same record separately for evidence production and relative-score disposition; it does not declare additional grader identities. The workflow annotation on that route adds that the circle-packing wrapper omits the overlap details from its ordinary explanation. This qualifies what reaches the coding agent without contradicting the runtime's task-dependent oracle or numerical checks. EPI-OBJ-attempt-feedback already retains task-specific omission limits. A result or evaluation-log pointer is not evidence of delivery of the exact violated constraint, successful diagnosis or recovery.

Pending observation is not grading failure. The workflow annotation on RT-RTE-eval-submission distinguishes CLI wait timeout from submission rejection and points to waiting again. The runtime record already separates pending return, terminal grading and polling timeout. Neither finding requires rollback or a new evaluation after a caller's wait expires.

Measured score is not knowledge acceptance or universal product admission. EPI-OBJ-attempt-score records metric-relative classifications; the first numeric attempt can be called improved. EPI-RTE-code-selection and WF-RTE-candidate-export can select candidates without a passing-score predicate. This complements the runtime distinction between committing a candidate, grading it and deploying a product. It does not conflict with score-driven scheduling in RT-RTE-heartbeat-feedback.

Structural review machinery and advisory semantic review coexist. MEM-CLM-grounding-gate, EPI-RTE-grounding-check and EPI-RTE-note-lint independently support the narrower scope of file/metadata checks and default advisory exits. EPI-RTE-synthesis-review adds a requested model textual-support review; EPI-RTE-confidence-disposition leaves action to the author. Memory explicitly left reviewer implementation uninspected, so the epistemic inspection adds coverage rather than refuting a claimed absence. Review-file existence, reviewer judgment, author acceptance and later integration remain separate.

Withdrawal is consumer-specific. MEM-RTE-note-maintenance establishes that the generated index excludes _archive/ while the hub walker can still enumerate archived Markdown. MEM-RTE-knowledge-query and EPI-RTE-note-selection agree that authored supersession/refutation metadata does not universally remove content. The discrepancy is between target readers, not between analyst findings. It prevents a synthesis claim of uniform epistemic retirement.

Recovery and relocation do not restore or erase all knowledge. Memory's annotations on RT-RTE-resume-steering and RT-RTE-island-migration preserve the runtime's non-atomic rollback and changed-neighborhood findings. Archived attempts can remain directly recoverable; notes and saved conversations are not thereby retracted. MEM-OBJ-session-state distinguishes session pointers from opaque conversation payloads. Shared-link changes do not establish removal of source-island information from a resumed conversation.

Delegation remains host-mediated. RT-RTE-specialist-delegation and its memory annotation agree on requested summaries and escaping shared-file edits, while leaving actual frame creation, trimming and budgets uninspected. MEM-RTE-note-maintenance is the distinct artifact-maintenance work, not another declaration of delegation transport. EPI-RTE-synthesis-review names a different specialist function, not a second librarian route.

Integration dispositions and limits

No unresolved disagreement between the supplied reports is identified. In particular, no dependent finding requires a supplied record to say something it no longer says. No required missing part was identified in the cited combined records: attempt score/feedback, migration output and research review already have declarations. No conflict is inferred merely from different inspection coverage, narrower scope or uncertainty.

The remaining memory integration issues stay local to their records: absent bundled optimizer helpers constrain MEM-OBJ-skills; flat metadata rewriting and uneven discovery constrain MEM-RTE-note-maintenance; CLI addressing, frontend fidelity and host ingestion constrain MEM-RTE-knowledge-query. The epistemic report's unassessed maintenance and skill evaluators do not negate these memory findings or establish a complete-system negative.

The workflow explicitly distinguishes its circle-packing task's docker-opencode configuration from the runtime's representative default Claude launch. It does not claim a successful execution of that exact configured path. RT-CLM-pi-launch concerns a separate manager/adapter interface mismatch, not a refutation of the default launch account. RT-CLM-isolation-scope retains conditional deployment protection rather than universal containment.

Finally, RT-CLM-agent-organizations, the memory revision account and the epistemic bounded conclusion distinguish orchestration, requested knowledge work and later benefit. None supplies observed instruction compliance, accepted novel knowledge, exercised theory-building or a causally demonstrated improvement from sharing. Their independent limits coexist with the implemented grading, scheduling and read-back positives; reconciliation selects no stronger combined status.